本文来源吾爱破解论坛
用到的知识
python基础语法
requests库
re正则表达式
xlwt 写入Excel库
os库用于写入文件目录
json 字符串转字典
import requests,os
import re
import xlwt
import time
import json
class get_yq_info:
def get_data_html(self):
headers = {'User-Agent': 'Mozilla/5.0 (Windows NT 6.1; WOW64) AppleWebKit/535.1 (KHTML, like Gecko) Chrome/14.0.835.163 Safari/535.1'}
response = requests.get('https://ncov.dxy.cn/ncovh5/view/pneumonia?from=timeline&isappinstalled=0', headers=headers, timeout=3)
# 请求页面
response = str(response.content, 'utf-8')
# 中文重新编码
return response
#返回了HTML数据
def get_data_dictype(self):
areas_type_dic_raw = re.findall('try { window.getAreaStat = (.*?)}catch\(e\)',self.get_data_html())
areas_type_dic = json.loads(areas_type_dic_raw[0])
return areas_type_dic
#返回经过json转换过的字典化的数据
def save_data_to_excle(self):
self.make_dir()
#调用方法检查数据目录是否存在,不存在则创建数据文件夹
count = 2
#数据写入行数记录
newworkbook = xlwt.Workbook()
worksheet = newworkbook.add_sheet('all_data')
# 打开工作簿,创建工作表
worksheet.write(1, 2, '省份名称')
worksheet.write(1, 3, '省份简称或城市名称')
worksheet.write(1, 4, '确诊人数')
worksheet.write(1, 5, '疑似人数')
worksheet.write(1, 6, '治愈人数')
worksheet.write(1, 7, '死亡人数')
worksheet.write(1, 8, '地区ID编码')
#写入数据列标签
for province_data in self.get_data_dictype():
provincename = province_data['provinceName']
provinceshortName = province_data['provinceShortName']
p_confirmedcount = province_data['confirmedCount']
p_suspectedcount = province_data['suspectedCount']
p_curedcount = province_data['curedCount']
p_deadcount = province_data['deadCount']
p_locationid = province_data['locationId']
#用循环获取省级以及该省以下城市的数据
worksheet.write(count, 2, provincename)
worksheet.write(count, 3, provinceshortName)
worksheet.write(count, 4, p_confirmedcount)
worksheet.write(count, 5, p_suspectedcount)
worksheet.write(count, 6, p_curedcount)
worksheet.write(count, 7, p_deadcount)
worksheet.write(count, 8, p_locationid)
#在工作表里写入省级数据
count += 1
#此处为写入行数累加,province部分循环
for citiy_data in province_data['cities']:
cityname = citiy_data['cityName']
c_confirmedcount = citiy_data['confirmedCount']
c_suspectedcount = citiy_data['suspectedCount']
c_curedcount = citiy_data['curedCount']
c_deadcount = citiy_data['deadCount']
c_locationid = citiy_data['locationId']
#该部分获取某个省下某城市的数据
worksheet.write(count, 3, cityname)
worksheet.write(count, 4, c_confirmedcount)
worksheet.write(count, 5, c_suspectedcount)
worksheet.write(count, 6, c_curedcount)
worksheet.write(count, 7, c_deadcount)
worksheet.write(count, 8, c_locationid)
#该部分在工作表里写入某城市的数据
count += 1
#此处为写入行数累加,cities部分循环
current_time = time.strftime("%Y-%m-%d-%H-%M-%S", time.localtime())
newworkbook.save('F:\人数采集3.0\实时采集v3.0-%s.xls' % (current_time))
print('======数据爬取成功======')
def make_dir(self):
file_path = 'F:/人数采集3.0/'
if not os.path.exists(file_path):
os.makedirs(file_path)
print('======数据文件夹不存在=======')
print('======数据文件夹创建成功======')
print('======创建目录为%s======'%(file_path))
else:
print('======数据保存在目录:%s======' % (file_path))
#检查并创建数据目录
def exe_task(self):
times = int(input('执行采集次数:'))
interval_time = round(float(input('每次执行间隔时间(分钟)')),1)
#round 方法保留一位小数
interval_time_min = interval_time * 60
for i in range(times):
get_yq_info().save_data_to_excle()
time.sleep(interval_time_min)
#执行完整采集任务
get_yq_info().exe_task()
版权声明:
本站所有资源均为站长或网友整理自互联网或站长购买自互联网,站长无法分辨资源版权出自何处,所以不承担任何版权以及其他问题带来的法律责任,如有侵权或者其他问题请联系站长删除!站长QQ754403226 谢谢。
- 上一篇: 借鉴网上教程,简单写了一个有声小说批量下载,菜鸟级别
- 下一篇: 无聊时做的本省疫情地图